网络结构
FCNVMB在代码层面更接近UNet结构,因此FCNVMB是UNet架构下的FWI。
在编码解码过程有如下三种操作:
- 红色箭头:特征重整和,不会发生尺码改变
- 紫色箭头:下采样,对特征进一步提取细化,尺寸缩小一半
- 黄色箭头:反卷积,导致尺寸扩大一倍
- skip connection:在编码过程中,首先利用上一次上采样 t t t的输出作为近端输入,再利用 t t t的对称的特征图,扩充过后作远端输入,将两个输入拼接,即可完成,如下是该过程实现代码
def forward(self, inputs1, inputs2):
'''
:param inputs1: Layer of the selected coding area via skip connection
:param inputs2: Current network layer based on network flows
:return:
'''
outputs2 = self.up(inputs2)
offset1 = (outputs2.size()[2] - inputs1.size()[2])
offset2 = (outputs2.size()[3] - inputs1.size()[3])
padding = [offset2 // 2, (offset2 + 1) // 2, offset1 // 2, (offset1 + 1) // 2]
# Skip and concatenate
outputs1 = F.pad(inputs1, padding)
return self.conv(torch.cat([outputs1, outputs2], 1))
这里的卷积层网络分为三个部分:
- nn.Conv2d二维卷积层
- nn.BatchNorm2d二维批归一化层
- nn.ReLU激活函数
if is_batchnorm:
self.conv1 = nn.Sequential(nn.Conv2d(in_size, out_size, 3, 1, 1),
nn.BatchNorm2d(out_size),
nn.ReLU(inplace=True), )
self.conv2 = nn.Sequential(nn.Conv2d(out_size, out_size, 3, 1, 1),
nn.BatchNorm2d(out_size),
nn.ReLU(inplace=True), )
else:
self.conv1 = nn.Sequential(nn.Conv2d(in_size, out_size, 3, 1, 1),
nn.ReLU(inplace=True), )
self.conv2 = nn.Sequential(nn.Conv2d(out_size, out_size, 3, 1, 1),
nn.ReLU(inplace=True), )
解码器
首先将输入的 ( 29 , 400 , 301 ) (29, 400, 301) (29,400,301)地震数据卷积为 ( 64 , 200 , 151 ) (64, 200, 151) (64,200,151),再特征重整和,最后下采样,重复此操作,直到变成 ( 1024 , 50 , 38 ) (1024, 50, 38) (1024,50,38),代码如下:
self.down1 = unetDown(self.in_channels, filters[0], self.is_batchnorm)
self.down2 = unetDown(filters[0], filters[1], self.is_batchnorm)
self.down3 = unetDown(filters[1], filters[2], self.is_batchnorm)
self.down4 = unetDown(filters[2], filters[3], self.is_batchnorm)
class unetDown(nn.Module):
def __init__(self, in_size, out_size, is_batchnorm):
'''
Downsampling Unit
[Affiliated with FCNVMB]
:param in_size: Number of channels of input
:param out_size: Number of channels of output
:param is_batchnorm: Whether to use BN
'''
super(unetDown, self).__init__()
self.conv = unetConv2(in_size, out_size, is_batchnorm)
self.down = nn.MaxPool2d(2, 2, ceil_mode=True)
def forward(self, inputs):
'''
:param inputs: Input Image
:return:
'''
outputs = self.conv(inputs)
outputs = self.down(outputs)
return outputs
解码器
解码器则是先将上次输出先进行上采样,再利用上采样结果作为近端输入,结合远端输入实现skip connection操作,如此循环重复
class unetUp(nn.Module):
def __init__(self, in_size, out_size, is_deconv):
'''
Upsampling Unit
[Affiliated with FCNVMB]
:param in_size: Number of channels of input
:param out_size: Number of channels of output
:param is_deconv: Whether to use deconvolution
'''
super(unetUp, self).__init__()
self.conv = unetConv2(in_size, out_size, True)
# Transposed convolution
if is_deconv:
self.up = nn.ConvTranspose2d(in_size, out_size, kernel_size=2, stride=2)
else:
self.up = nn.UpsamplingBilinear2d(scale_factor=2)
def forward(self, inputs1, inputs2):
'''
:param inputs1: Layer of the selected coding area via skip connection
:param inputs2: Current network layer based on network flows
:return:
'''
outputs2 = self.up(inputs2)
offset1 = (outputs2.size()[2] - inputs1.size()[2])
offset2 = (outputs2.size()[3] - inputs1.size()[3])
padding = [offset2 // 2, (offset2 + 1) // 2, offset1 // 2, (offset1 + 1) // 2]
# Skip and concatenate
outputs1 = F.pad(inputs1, padding)
return self.conv(torch.cat([outputs1, outputs2], 1))
self.up4 = unetUp(filters[4], filters[3], self.is_deconv)
self.up3 = unetUp(filters[3], filters[2], self.is_deconv)
self.up2 = unetUp(filters[2], filters[1], self.is_deconv)
self.up1 = unetUp(filters[1], filters[0], self.is_deconv)
self.final = nn.Conv2d(filters[0], self.n_classes, 1)
预训练
可以先使用模拟数据进行预训练,最后再使用真实数据训练得到最终的Re-trained model
与InversionNet的异同点
相同点:
- 无论是FCNVMB还是InversionNet, 它们都是单一的端到端深度网络
并没有利用更多的物理含义. - 都采用了编码器-解码器的架构.
- 都是利用叠前多炮数据的不同炮集直接投入训练, 并未处理.
不同点:
- InversionNet在编码的过程中最终将图像压缩为完全的一维向量, 抛弃了空间关联性; 而FCNVMB在压缩后仍保留了25 * 19的空间尺寸关联. 但是对应的代价是FCMVMB的模型内存占有量很大, 难以进行大batch体量的训练; 而InversionNet可以非常灵活进行训练.
- FCNVMB面向SEG数据, InversionNet面向部分OpenFWI的数据. 因为OpenFWI数据的特点, InversionNet有非常明显的高度降维部分.
- FCNVMB使用了迁移学习的训练手段, 后者InversionNet是单一的训练思想.
- FCNVMB采用了包含skip connection的UNet的架构, 而InversionNet是单一的CNN架构.